How much capital does an insurer writing group life and property business need for the two lines together, and how much does the answer depend on the copula used to join them? This project builds a true loss model in which the lines depend on each other through shared shocks, fits the standard copula families (independence, Gaussian, Student t, Gumbel) to histories drawn from it, and compares the capital each gives with the true capital.
Because the true model is known, every fitted capital figure can be checked against the right answer, which is not possible with real data.
Write-up: https://mdingi.co.za/work/cross-line-tail-risk
All amounts are annual claims in R million on synthetic data. Capital is the one-year 99.5% VaR less the expected claims, the calibration standard used in Solvency II-style regimes.
From reports/results.md (run.py, seed 123):
The true model. Expected claims of 2,783 a year. Capital of 800 for life alone and 1,282 for property alone, but 1,798 for the two together: a diversification benefit of 13.6%. 78% of the years beyond the combined 99.5% VaR are pandemic years.
The copula alone. Each family fitted to 200,000 true years and combined with the true marginals, so the only error is the shape of the dependence:
| Copula | Capital | Against truth | Diversification benefit |
|---|---|---|---|
| Independence | 1,387 | -22.9% | 33.4% |
| Gaussian | 1,485 | -17.4% | 28.6% |
| Student t | 1,506 | -16.2% | 27.4% |
| Gumbel | 1,600 | -11.0% | 22.9% |
| True model | 1,798 | 13.6% |
All four understate the capital. Kendall's tau is only 0.12, so a copula calibrated to the body of the data sees weak dependence, but in the true model the lines move together most in the worst years: given that life claims are above their 99th percentile, property claims are above theirs 31% of the time. The Gaussian copula gives 3.9% and the Gumbel 18.1%. AIC correctly prefers the Gumbel, the only family with upper tail dependence, and it is still 11% short.
The whole model fitted to a history. In practice the marginals are estimated too. Over 500 simulated 40-year histories, the capital from the AIC-chosen copula has a median of 1,270 (5th to 95th percentile 676 to 3,358) and is below the truth 74% of the time. With 200 years of history the median is 1,445 and it is below the truth 85% of the time: more data narrows the range around the wrong answer. The main cause is the life marginal: a lognormal or gamma fitted to annual totals cannot carry a rare pandemic tail, so its median capital is 252 from 40 years and 309 from 200 years, against a true 800. With the true marginals, the fitted copulas' error is the 11% to 23% above.
The observed 40-year history happens to contain no pandemic (as 38% of 40-year histories do). Fitted to it, every copula gives capital of 901 to 953, about half the truth, and AIC picks independence.
src/truth.py, with every parameter in src/config.py:
- Life: base death claims, lognormal with mean 1,500 and log standard deviation 0.05.
- Property: attritional claims, lognormal with mean 1,100 and log standard deviation 0.06, plus catastrophe events at 0.8 a year whose losses above the reporting threshold of 75 are generalised Pareto (shape 0.3, scale 80).
- Everyday dependence: the two lines' log base claims have correlation 0.25 (claims inflation and the economy).
- Floods: 40% of catastrophe events are floods, which also cause death claims of about 3% of the property loss.
- Pandemics: one year in 40 on average. Death claims rise by a lognormal share of base claims (median 35%), and business-interruption claims of about 1,500 times that share (lognormal noise, log standard deviation 0.4) are recorded as one property event.
The parameters are illustrative, chosen so that the two lines need comparable capital on their own; they are not estimates from any real insurer.
- Marginals. Life: lognormal and gamma fitted by maximum likelihood, chosen by AIC. Property: attritional claims (annual property claims less the recorded events) lognormal; event frequency Poisson; event losses above 75 generalised Pareto by maximum likelihood. The annual property distribution is simulated from these, 1,000,000 years.
- Copulas. Fitted to the pseudo-observations (ranks over n + 1) by maximum pseudo-likelihood, so the copula does not depend on the marginal fits. Selection by AIC.
- Capital. 1,000,000 years simulated from each fitted copula and the marginals. Capital is VaR 99.5% less the mean, for each line alone and combined; the diversification benefit is 1 - combined capital / sum of the standalone capitals.
- Truth and best fits. The true capital from 4,000,000 years of the true model. Each family's best fit is fitted to 200,000 true years and combined with the true marginals.
- Repeated histories. 500 histories each of 40 and 200 years, refitting everything and recording each family's capital with the fitted and with the true marginals (200,000 simulated years per fit, 8 processes).
reports/methodology.md has the copula densities, samplers and tail formulas.
pip install -r requirements.txt
python3 -m pytest -q tests # 18 tests, about a minute
python3 run.py # about 5 minutes on 8 coresrun.py writes output/results.json, output/study.csv (every replication),
output/history.csv and output/history_events.csv (the observed history), the figures in
output/figures/ and reports/results.md.
The tests check that each copula's density integrates to 1, that the samplers have uniform margins and the right Kendall's tau and tail dependence, that the fits recover known parameters, that the true model's means match their formulas, and that the marginal fits and capital measures are correct on known cases.
- Synthetic data with illustrative parameters.
- One-year horizon, no reserve risk, no reinsurance, and the history is assumed already restated to today's exposure.
- Only one-parameter and two-parameter copulas fitted to annual totals. The natural remedy for what this study finds, modelling the pandemic as an explicit common shock with its frequency and severity set by scenario judgement, is not implemented.
- Monte Carlo error: about 1% of the capital at 200,000 simulated years (the spread of the independence capital with true marginals across replications, which is pure simulation noise), so about 0.5% at 1,000,000.
src/config.py settings and the true model's parameters
src/truth.py the true loss model
src/copulas.py copula densities, fits, samplers, tail dependence
src/marginals.py life and property marginal fits
src/capital.py VaR, TVaR, capital and diversification
src/study.py repeated-history study
src/figures.py figures
run.py runs everything
tests/ pytest suite
reports/ methodology.md, results.md
output/ results, history, study and figures
Sbusiso Mdingi


